Edinburgh Research Explorer

Transcription

1 Edinburgh Research Explorer Personality traits below facets Citation for published version: Mõttus, R, Kandler, C, Bleidorn, W, Riemann, R & McCrae, RR 2016, 'Personality traits below facets: The consensual validity, longitudinal stability, heritability, and utility of personality nuances' Journal of Personality and Social Psychology., /pspp Digital Object Identifier (DOI): /pspp Link: Link to publication record in Edinburgh Research Explorer Document Version: Peer reviewed version Published In: Journal of Personality and Social Psychology General rights Copyright for the publications made accessible via the Edinburgh Research Explorer is retained by the author(s) and / or other copyright owners and it is a condition of accessing these publications that users recognise and abide by the legal requirements associated with these rights. Take down policy The University of Edinburgh has made every reasonable effort to ensure that Edinburgh Research Explorer content complies with UK legislation. If you believe that the public display of this file breaches copyright please contact openaccess@ed.ac.uk providing details, and we will remove access to the work immediately and investigate your claim. Download date: 28. May. 2016

2 RUNNING HEAD: Personality characteristics below facets Personality Traits Below Facets: The Consensual Validity, Longitudinal Stability, Heritability, and Utility of Personality Nuances Date of submission: 2 nd March

3 Abstract It has been argued that facets do not represent the bottom of the personality hierarchy even more specific personality characteristics, nuances, could be useful for describing and understanding individuals and their differences. Combining two samples of German twins, we assessed the consensual validity (correlations across different observers), rank-order stability, and heritability of nuances. Personality nuances were operationalized as the 240 items of the Revised NEO Personality Inventory (NEO-PI-R). Their attributes were examined by analyzing item residuals, controlling for the variance of the facet the item had been assigned to and all other facets. Most nuances demonstrated significant (p <.0002) cross-method agreement and rank-order stability. A substantial proportion of them (48% in self-reports, 20% in informant ratings, and 50% in combined ratings) demonstrated a significant (p <.0002) component of additive genetic variance, whereas evidence for environmental influences shared by twins was modest. Applying a procedure to estimate stability and heritability of true scores of item residuals yielded estimates comparable to those of higher-order personality traits, with median estimates of rank-order stability and heritability being.77 and.52, respectively. Few nuances demonstrated robust associations with age and gender, but many showed incremental, conceptually meaningful, and replicable (across methods and/or samples) predictive validity for a range of interest domains and body mass index. We argue that these narrow personality characteristics constitute a valid level of the personality hierarchy. They may be especially useful for providing a deep and contextualized description of the individual, but also for the prediction of specific outcomes. 2

4 Keywords: personality hierarchy; nuances; heritability; stability; prediction Personality characteristics below facets 3

5 Personality Traits Below Facets: The Consensual Validity, Longitudinal Stability, Heritability, and Utility of Personality Nuances The characteristics often used to describe human personality variation are traits: relatively consistent patterns of emotion, cognition, and behavior in which individuals differ from one another. Traits are thought to be arranged hierarchically. At the highest level, personality trait variation can be described along only a few overarching dimensions. For example, one- (Rushton, Bons, & Hur, 2008) and two- (DeYoung, 2006) factor levels of the hierarchy have been proposed (but also challenged; see Ashton, Lee, Goldberg, & de Vries, 2009; McCrae et al., 2008). The most widely endorsed model of broad traits, however, is the Five-Factor Model of personality (FFM; McCrae & John, 1992) or the Big Five (Goldberg, 1990). At lower levels, the broad traits are thought to split into increasingly narrow constructs. For example, each FFM domain can be split into two aspects (DeYoung, Quilty, & Peterson, 2007) or more numerous facets (Costa & McCrae, 1992). These different levels of generality/specificity reflect a trade-off. Like all sciences, personality research strives for simplicity and parsimony, which is why researchers sometimes prefer to describe and, perhaps, explain personality variation using as few and comprehensive constructs as possible. And indeed, this has proven a useful strategy. For example, broad traits such as those of the FFM can be used to map personality variability across demographic and cultural groups (e.g., Bleidorn et al., 2013; Costa, Terracciano, & McCrae, 2001; Rentfrow et al., 2013; Schmitt, Realo, Voracek, & Allik, 2008; Terracciano, McCrae, Brant, & Costa, 2005) and predict important life outcomes such as occupational 4

6 success, divorce, and longevity (Ozer & Benet-Martínez, 2006; Roberts, Kuncel, Shiner, Caspi, & Goldberg, 2007). They are systematically linked to several personality-relevant characteristics, such as self-esteem, control beliefs, motives, values, interests, attitudes, and subjective well-being (Kandler, Zimmermann, & McAdams, 2014). Likewise, they provide a framework for diagnostic distinctions in psychopathology (Krueger & Markon, 2014). On the other hand, science also strives for comprehensiveness and precision, and therefore using a few broad personality dimensions may not always be optimal. Narrower, lower-order traits such as aspects and facets can often provide useful incremental information for the purpose at hand. Facets have been shown to outperform broad traits in the prediction of specific behaviors (Paunonen & Ashton, 2001), general job performance (Judge, Rodell, Klinger, Simon, & Crawford, 2013), and personality disorder diagnoses (Reynolds & Clark, 2001). The specific variance in facet scores assessed by statistically controlling for variance from the five broad FFM domains has been shown to be consensually valid (agreed upon by different observers; Kandler, Riemann, Spinath, & Angleitner, 2010; Costa & McCrae, 2008; Mõttus et al., 2014) and genetically influenced (Briley & Tucker-Drob, 2012; Jang, McCrae, Angleitner, Riemann, & Livesley, 1998). This evidence supports the hypothesis that the specific variance reflects something substantive about personality. Facet-level traits are therefore not simply interchangeable operationalizations of the broad factor they were designed to reflect or blends of two or more factors (Hofstee, DeRaad, & Goldberg, 1992). Instead, they are traits in their own right and make unique contributions to the depiction of personality. In this article, we examine whether the same conclusions can be drawn for characteristics at an even 5

7 lower level of the personality hierarchy, using cross-sectional and longitudinal multi-method data from twins. Nuances as Traits McCrae (2015) proposed that there is a meaningful level of the trait hierarchy below facets, called nuances, which correspond roughly to single items (or groups of very similar items) in a facet scale. For example, bitterness and touchiness may be different nuances of angry hostility, a facet of Neuroticism. Nuances may specify either the eliciting situation (e.g., fear of heights as a source of anxiety or inability to accept criticism as source of anger) or the characteristic response to a range of situations (e.g., a nervous tic as an expression of anxiety across different circumstances or feeling offended as a result of criticism of any kind). Initially, nuances were evoked to account for the fact that retest reliability is generally a better predictor of facet scales differential stability, heritability, and consensual validity than is internal consistency (McCrae, Kurtz, Yamagata, & Terracciano, 2011). McCrae (2015) argued that this occurs because retest reliability reflects both the variance common to items in a facet scale and the item-specific variance that distinguishes different nuances of the same facet. Here we seek to directly investigate the properties of nuances and establish at least some of them as substantive personality characteristics. A number of psychometricians have acknowledged that facet-level traits are composed of more specific elements, but it has not been clear whether these elements are themselves lower-level traits. Eysenck (1967, 1991) argued that traits are composed of habitual responses, and habits are usually 6

8 distinguished from traits. Jackson (1971) regarded scale items as manifestations of the trait that covered the gamut of situations and modes of response appropriate for sampling broadly a personality construct (p. 234), a formulation that suggests that items are merely samples of the broader trait, not traits in their own right. Goldberg (1993) recognized traits at many levels of the hierarchy, but he conceived of traits as phenotypic descriptors that might or might not be temporally stable and may not reflect underlying causal influences (Saucier & Goldberg, 1996) views far removed from the classic formulation of traits as enduring underlying dispositions. Also, some researchers have conceived of traits at higher levels of personality hierarchy as emergent from direct associations between specific characteristics that could potentially be represented by single items (Cramer et al., 2012) or from the direct links of these characteristics with beliefs and motivational constructs (Wood, Hensler Gardner, & Harms, 2015). Five-Factor Theory (FFT; McCrae & Costa, 1999) states that traits are basic tendencies, organized hierarchically from narrow and specific to broad and general dispositions (p. 145), but it does not specify how many levels the trait hierarchy has, or what lower limit (if any) it has. Instead, it specifies definitional criteria for traits. Specifically, FFT holds that traits are not groundless attributions, but enduring dispositions, and adds that they are also genetically based 1. This conceptualization provides criteria for determining whether nuances are traits: If so, they must be detectable by different observers, stable over substantial periods of time, and have some demonstrable 1 FFT requires some biological basis for traits, but not necessarily a genetic one: diet, disease, or drugs might also affect traits. However, with the present data we are only able to address genetic influences. 7

9 genetic foundation. Evidence that nuances meet these criteria would expand what have typically been considered traits in FFT-based research from domains and facets to narrower characteristics. Individual items are likely to show some degree of consensual validity, stability, and heritability, if only because they share variance with, or reflect, facets known to have those properties. However, the claim that nuances constitute a distinct level of the trait hierarchy requires evidence that they contribute something unique to the characterization of the individual evidence of the consensual validity, stability, and heritability of that portion of the items variance that is not shared with facets. Also, the unique variance in single items could demonstrate associations with external variables (i.e., variables not explicitly included in personality frameworks such as FFM domains and their facets). Extending the personality hierarchy below facets to nuances could have important conceptual and practical implications. Conceptually, this would refine our understanding of the phenotypic architecture of traits (McCrae, 2015). For example, if there are nuances that do not appear to be merely interchangeable reflections of ostensible higher-level domains or facets, then this would suggest that domains and even facets aggregate numerous specific etiological mechanisms. Empirical and theoretical work on this hypothesis could lead to refinement of trait theories such as the FFT. Practically, nuances might have unique links with etiological factors such as genetic variants or brain parameters, offer incremental validity in the prediction of important outcomes, and have implications for test construction and data analytic practices. For example, tests that incorporate different sets of nuances of facets could not be expected to measure exactly the same facets, which has implications for how their links with other variables can be interpreted (Mõttus, 2015). 8

10 Cross-Method Agreement on Nuances The first direct evidence for the proposal that nuances have trait-like characteristics was provided in a study by Mõttus, McCrae, Allik, and Realo (2014). They examined data from the Estonian version of the NEO Personality Inventory-3 (NEO-PI-3; McCrae & Costa, 2010) and found that item residual scores from which the common variance of the facet had been removed showed significant crossrater (i.e., cross-method) agreement. Individual items are likely to include substantial error variance; as a result, the magnitude of agreement on item-specific variance was smaller than that typically found in studies of domains and facets (Mõttus et al., 2014). The mean correlation between raters on individual items was.31; when item residuals were analyzed, this mean agreement was reduced to.19. However, statistically significant agreement was found for most of the 240 residualized items (range =.06 to.47). This suggests that the different ways in which the same facet can be expressed (or, alternatively, their more or less autonomous constituents) are real in the sense of being detectable by different raters. In the present study, we attempted to replicate this finding with German multi-method data. Stability of Nuances Cross-method agreement would appear to be a necessary condition for considering nuances to be traits, but it is not sufficient. Small variations in the ways facet-level traits are expressed might be transient, perhaps reflecting current situational pressures that are visible to all observers. For example, a person who is normally high in the activity facet of Extraversion may have suffered a recent physical injury which temporarily alters his or her response to the item I act forcefully and energetically; the 9

11 response of an informant to this item may also be temporarily altered. If nuances are traits, they and their specific variance should show substantial rank-order stability over an interval of years. To the best of our knowledge, this has not yet been investigated. In this article, therefore, we examined whether nuances also show rank-order stability, a fundamental requirement for considering them traits. Genetic and Environmental Variance in Nuances Nuances as enduring dispositions might be considered traits, but it is possible that they have a distinct type of etiology from facets and domains. Higher order traits are known to be substantially heritable (Bleidorn, Kandler, & Caspi, 2014), and in some theories (e.g., Five-Factor Theory; McCrae & Costa, 1999) traits are defined as dispositions with a heritable basis. Similarly, Turkheimer, Pettersson, and Horn (2014) asserted that all traits are heritable (p. 532). From this perspective, the question is: Are nuances really traits? It is possible that the distinguishing features of nuances are entirely learned, as experience shapes the expression of a given facet. That is, the mechanisms that make nuances co-vary and thereby give rise to higher-order traits may be heritable, whereas influences that make them different may reflect purely environmental effects (cf. Borkenau, Riemann, Angleitner, & Spinath, 2001). If this were true, nuances would be qualitatively different from facets and domains; by the standards of the FFT, they would be characteristic adaptations (McCrae & Costa, 1999). To examine this issue, we conducted behavioral genetic analyses of items and their specific variance. Evidence of substantial heritability would suggest that nuances are qualitatively similar to higher-level traits. Furthermore, evidence for either genetic or shared environmental effects (i.e., shared by family members such as twins reared together) on the specific variance of items would provide further 10

12 evidence that nuances reflect 'signal' (some kind of substantive variance) rather than mere 'noise' (artifact or random error). Demographic Variations in Nuances Nuances could provide incremental value for our understanding of personality variability across demographic groups. For example, Mõttus and colleagues (2015) showed that a substantial proportion of age-group differences in personality characteristics could be ascribed to unique variance of single personality test items. None of the 30 NEO-PI-3 facet scales used in the study met the criteria for strong (scalar) measurement invariance across age groups, suggesting that items of the same scales varied in age-trajectories. Furthermore, 46% of items had significant (p <.0002, i.e., after Bonferroni correction) correlations with age when residualized for the variance of their respective facet scores. To the best of our knowledge, similar analyses have not yet been carried out for gender. In this paper, we will examine the extent to which residual variance in single items is linked with age and gender. Predictive Utility of Nuances Personality characteristics predict important outcomes, such as subjective wellbeing, occupational success, divorce, and even mortality (Roberts et al., 2007). Hence, the case for nuances as personality traits would be strengthened by showing that they also predict incremental variance in outcomes (although we believe that nuances need not predict outcomes in order to broaden our conceptual understanding of how personality is organized). To date, there is very little evidence that nuances predict outcomes but there is an obvious reason for such paucity of evidence: With some 11

13 notable exceptions (e.g., Buss, Block, & Block, 1980; Kolar, Funder, & Colvin, 1996) researchers almost never examine individual items as predictors of outcomes. A full demonstration of the utility of examining nuances as predictors of outcomes would require a lengthy program of research, examining a wide range of outcomes, both broad and narrow. Picking one or a few outcomes and finding these to have limited associations with nuances would not be a conclusive test, because not every nuance is likely to have relevance for every outcome. Broad traits at least their operationalizations as personality test scores summarize wide ranges of behaviors and broad outcomes also summarize the cumulative effects of multiple behaviors, which makes some associations between them almost inevitable (Mõttus, 2015). Nuances do not summarize wide ranges of behaviors and are, for that reason alone, likely to have fewer links with outcomes, either broad or specific. They might, however, offer some small incremental validity in the prediction of broad criteria, and if these criteria were of great importance such as happiness or mortality even a small contribution would be welcome. Perhaps more importantly, nuances might have substantial utility in the prediction of narrow criteria to which they are directly relevant. Because personality inventories typically include many items, an unstructured search for correlates would likely yield many false positive associations. For broad criteria, systematic exploration of the full set of item predictors might be reasonable if large samples are used, stringent significance levels are chosen, and, most importantly, replication is required before reaching conclusions. For narrow criteria, it may be more reasonable to test specific a priori hypotheses for a selected subset of items (although replication is important in this case, too). This is similar to the two strategies for 12

14 identifying the molecular genetic correlates of phenotypic traits: genome-wide and candidate gene approaches (Van Gestell & Van Broeckhoven, 2003). In this article, we illustrate these kinds of research by exploring the value of individual items in predicting the broad criteria of conservative attitudes (typically correlated with openness and conscientiousness; Jost, Federico, & Napier, 2009) and satisfaction with life (typically associated with variation in neuroticism, extraversion, agreeableness, and conscientiousness; Steel, Schmidt, & Shultz, 2008). We also employed the second strategy of rationally matching specific nuances with specific outcomes for the narrower criteria of body mass index (BMI) and interests in various domains of life. The Analysis of Single Items Analyses of single items and item residuals pose special challenges. Item scores may be less reliable than aggregate scale scores, so effect sizes are likely to be small, and large samples may therefore be needed for appropriate statistical power. Individual items are also subject to the same artifacts of method (e.g., evaluative bias, extreme responding) as longer scales, so multi-method analyses may be advisable. Likewise, item responses are often based on an ordinal scale with a limited number of possible values and the responses are often strongly skewed, both of which artificially constrain the available variance. Of particular importance is that all single items are inherently unbalanced scales, potentially distorted by acquiescent responding. We address these concerns below. 13

15 Rank-Order Stability of Nuances There is a large literature on the stability of domains (Roberts & DelVecchio, 2000; Terracciano, Costa, & McCrae, 2006). In general, stability increases with age, and by mid-adulthood, traits typically show stability coefficients ranging from.60 to.90 over intervals as long as 10 or more years (Roberts & DelVecchio, 2000). Stability coefficients of this magnitude can be found in both self-report and informant rating data (Costa & McCrae, 1992). They are, however, probably underestimates of true stability, because the observed stability coefficients are attenuated by retest unreliability. When corrected for unreliability and method-specific variance, stability coefficients are often greater than.90 (Kandler, Bleidorn, Riemann, Spinath, Thiel, & Angleitner, 2010). In this study, we examined self-reports and aggregated informant ratings of twin pairs assessed on two occasions five years apart using the German version of the NEO Personality Inventory (NEO-PI-R; Costa & McCrae, 1992; Ostendorf & Angleitner, 2004). Long-term retest correlations for single items represent the stability of nuances, but in principle that stability might be attributable solely to the common variance: Perhaps each item is stable simply because it assesses a stable facet. Therefore, we also present more informative analyses based on item residuals, from which the common facet variance has been removed. McCrae (2015) offered an analysis of the typical variance components in single NEO Inventory items; not surprisingly, most (51%) of the variance was attributable to random error. Consequently, one should expect that the average observed stability of single items would not exceed about.50. Item residuals consist of specific variance plus random error, and McCrae s (2015) analysis suggests that 14

16 about two-thirds of this total is error. We would thus expect the observed stability of item residuals to be no higher than.33. Ideally, a comparison of the stability of factors, facets, and nuances would be based on true scores, or on stability coefficients disattenuated for retest unreliability. Data on the retest reliability of individual NEO items and item residuals are not yet available, but it is possible to estimate true score stability by using multi-method longitudinal data. For example, if one calculates the correlation between self-reports of a characteristic at the first measurement occasion and informant ratings of the same characteristic at the second measurement occasion, the resulting stability coefficient should be relatively free from both self- and informant-report method biases, because these biases are not likely to be shared. However, this correlation would be attenuated by imperfect agreement between self- and informant-reports even when collected at the same time. The concurrent cross-method correlation thus represents an upper limit to observed cross-method stability. It then follows that true score stability can be estimated as the ratio of the cross-lagged, cross-method correlation to this upper limit (McCrae, 1994). This is similar to the widely-used procedure of correcting a correlation for unreliability, if we think of cross-rater agreement as a reliability estimate. We used this method to estimate the true-score stability of items and item residuals. Genetic and Environmental Variance in Nuances We applied standard behavioral genetic analyses to item scores and their residuals to estimates the genetic and environmental sources of variance in nuances. It is possible to partition individual differences in item scores and item residuals into components that estimate additive genetic effects (a²), 15

17 genetic dominance effects within gene loci (d²), epistatic gene gene interaction effects between gene loci (b²), twins shared environmental effects (c²), and environmental effects not shared by twins reared together, including random error variance (e²). Whereas variance due to additive genetic effects reflects narrow-sense heritability, variance due to both additive and nonadditive genetic influences (such as dominance or epistatic effects) reflects broad-sense heritability. However, with a simple twin design, environmental effects shared by twins and nonadditive genetic influences cannot be estimated in the presence of each other. Therefore, only three of the five effects can be estimated in a single model. The usual procedure in such cases is to examine different models (σ² = a² + c² + e², σ² = a² + d² + e², and σ² = a² + b² + e²) and select the best-fitting model. Our data allowed us to conduct such analyses with twins self-reports and with averaged informant ratings. A comparison of results would show how robust the findings are across different methods of measurement, which is an interesting question in its own right. Additionally, we also carried out the analyses in combined self- and informant-reports. It is also possible to estimate the true score heritability of item residuals using a procedure similar to that employed to estimate the longitudinal stability of true scores. In this analysis, the disattenuated, random and nonrandom error-reduced estimate of the similarity of twins was calculated as the ratio of the cross-twin, cross-informant correlation to the within-twin, cross-informant correlation. For example, one can correlate Twin A s self-report with informant ratings of Twin B, and divide this by the correlation between Twin A s self-report and informant ratings of Twin A. This estimate can be calculated for both monozygotic (MZ) and dizygotic (DZ) twins, and standard behavioral genetic formulas can then be used to estimate a², d², b², c², and e² components. These estimates may provide a 16

18 rough indication of genetic and environmental influences on individual differences in true scores of nuances. Estimates of the stability and heritability of true scores are derived from a series of calculations using item data that is intrinsically less reliable than aggregated data. Thus, the estimate obtained for any given item is likely to have a very wide margin of error. Our focus, however, was on the distribution of estimates and not on any single one of them; the median and interquartile range are probably relatively robust, providing a sense of the stability or heritability of the true score for a typical nuance. Recall, however, that these are ideal values, essentially the upper limit of what might be observed if very reliable measures of nuances, based on ratings of many items made by many observers, were analyzed. Therefore, the estimates that we provide are of more theoretical than practical value. Variations Across Items Although operationalizing nuances as single items is a reasonable way to begin the study of this level of the trait hierarchy, it is unlikely that each of the 240 NEO Inventory items corresponds to a unique nuance. Two or more items might tap the same nuance a likely possibility, given that Mõttus and colleagues (2014) found residual correlations between many items after controlling for factor and facet variance. Some items might be relatively pure measures of the facet they reflect, with little specific variance. Other items, however, may have a more substantial specific component, and it would be useful to identify these items, because these are the items most likely to add incremental validity in 17

19 the prediction of outcomes or otherwise contribute to our understanding of personality differences and their development. We hypothesized that the cross-observer agreement, longitudinal stability, and genetic component of item residuals would be larger for items with high saturations of nuance-specific variance beyond random error and method variance. If so, observed self-informant agreement, rank-order stability, and heritability estimates should have been positively intercorrelated across the set of 240 items. Such regularity would support the claim that there is substantive residual variance in particular items. Items could then be ranked by combining these three indicators of the quantity of specific variance. We examined these rankings to identify and characterize the items with most nuance-specific variance. Facets, too, will vary in the amount of item-specific variance they contain. McCrae (2015) estimated these values by subtracting the internal consistency from the retest reliability of each of the 30 NEO Inventory facets, using American and international data. In the present German data, it is possible to make a corresponding estimate by calculating the mean cross-observer agreement, longitudinal stability, and heritability of item residuals across the eight items in each facet. Correlations of these means with the estimates provided by McCrae allowed for a direct test of the hypothesis that retest reliability exceeds internal consistency because it includes item-specific variance. 18

20 Method Participants Cross-sectional sample. Combined data from the third wave of the Bielefeld Longitudinal Study of Adult Twins (BiLSAT; Kandler, Riemann, Spinath, Bleidorn, Thiel, & Angleitner, 2013) and from the Jena Twin Study of Social Attitudes (JeTSSA; Stößel, Kämpfe, & Riemann, 2006) were used. Only participants with at least two sets of personality ratings (either data for both twins and/or for self- and informant-reports) were used in this study. The resulting sample consisted of 1599 individuals (1231 females), who were between 17 and 82 years old (M = 36.02; SD = 13.41). Among these participants were 690 complete twin pairs: 432 monozygotic (MZ) pairs (74 male pairs and 358 female pairs) and 258 dizygotic (DZ) pairs (34 male, 134 female, and 90 opposite-sex pairs). Both self- and informant ratings were available for 1,492 people (93%). Among the twins, informant ratings were not available for 107 people; for 26 twin pairs, informant ratings were not available for either pair-member. For genetic analyses that required informant ratings to be available for both siblings of a twin pair, data from 609 twin pairs (88%) could be used. Most participants were instructed to ask acquaintances, who knew them well but preferably did not know their twin sibling, to provide the informant ratings. Thus, there were different, quasi-independent informants for twin siblings. Most informants were friends and spouses. Reports from multiple informants were averaged when available (in 88% of cases). Longitudinal sample. For longitudinal analyses, data from the third and fourth waves of BiLSAT was used (N = 400; 337 females). In particular, the sample consisted of 200 twin pairs tested twice over a period of about five years. At wave three, participants' ages ranged from 22 to 74 with a mean of 19

21 41.47 (SD = 13.80), whereas at wave four the average age was years (SD = 13.81). Among the twins were 138 MZ pairs (12 male pairs and 126 female pairs) and 62 DZ pairs (12 male, 35 female and 15 opposite-sex pairs). For all twins a pair of informant ratings was available at each measurement occasion, although the informants were not always the same on both occasions. Estonian Genome Bank. The associations between BMI and item residual variances were replicated in the data published by Vainik et al. (2015; N = 2,581) with 982 added cases (total N = 3,563; age range 18 to 91 years with a mean of and SD of 17.01; 1,434 men). In these data, both self- and informant-ratings of personality were available (for details see Vainik et al., 2015). Measures Personality nuances. For the measurement of personality characteristics, the German version of the NEO-PI-R was administered (Costa & McCrae, 1992; Ostendorf & Angleitner, 2004). The NEO-PI- R contains 240 items, grouped into 30 facet scales, which are hierarchically organized under the five domain scales of the FFM. Responses were made on a 5-point Likert scale ranging from strongly disagree to strongly agree. The domains of the NEO Inventory correspond closely to the five factors found in analyses of many personality inventories (see Markon, Krueger, & Watson, 2005), and its facets were chosen to represent important traits within each domain. For these reasons, the NEO Inventory item pool seems to provide a broad sample of nuances; clearly, however, it does not exhaust the population of nuances. In the Estonian Genome Bank data, personality ratings were obtained using the Estonian translation of the NEO-PI-3 (McCrae & Costa, 2010), which is a slightly modified version of the NEO-PI-R. 20

22 Conservatism. At the third wave of BiLSAT, twins self-reported political attitudes were captured with a German 35-item version of the Wilson-Patterson C Scale (Schiebel, Riemann & Mummendey, 1984; see also Riemann, Grubich, Hempel, Mergl, & Richter, 1993). This measure was actualized with respect to catchphrases including current political topics (e.g., promotion of alternative energy resources, homosexual marriage, or German contribution to UN). Responses were made on a 7- point Likert scale ranging from strongly disagree to strongly agree. Items endorsing liberal positions were reverse coded. Internal consistency was good (α =.86). Life satisfaction. At the fourth wave of BiLSAT, twins completed the German version of the Satisfaction With Life Scale (Diener, Emmons, Larson, & Griffin, 1985; Glaesmer, Grande, Braehler, & Roth, 2011). Responses were made on a 7-point Likert scale ranging from strongly disagree to strongly agree. Internal consistency was good (α =.86). Interests. Interests were measured with the German General Interest Scale (GIS; Brickenkamp, 1990), a questionnaire including 48 unipolar items developed to map the intensity and breadth of individual interests across diverse fields of interest covering music, computer technology, arts, science, architecture, biology, literature, nutrition, politics, agriculture, business, fashion, education, sport, medicine, and entertainment. Each field is represented by three items that reflect different types of activities in the specific field of interest: (a) consuming and receptive activities (e.g., watching sports), (b) producing, participating, and imitating activities (e.g., getting exercise), and (c) creative and inventive activities (e.g., developing new methods of training). Both self-reports and informant-ratings were available. 21

23 Body mass index (BMI). In the German twin data, self-reported height and weight were used. The weight and height of the members of the Estonian Genome Bank were measured by trained medical staff at the recruitment. BMI was calculated as the ratio of weight (kg) to height (m 2 ). All these data were collected in a manner consistent with ethical standards for the treatment of human subjects. Results Cross-Method Agreement on Nuances Cross-method agreement on raw item scores and item residuals was estimated in the crosssectional sample by pooling data from all individuals for whom self- and informant-reports were available (N = 1,492). Items were residualized for their respective facet scores (with the item itself excluded) and scores of all other facets using linear regression. Note that this procedure also means that items were residualized for the FFM domains, because these are linear combinations of facets. Here and henceforth, unless otherwise noted, we used a conservative Bonferroni-corrected p-value threshold of.0002 for testing the significance of item-level associations to minimize type 1 error rate. Across the 240 items, self-informant correlations for raw (unresidualized) item scores ranged from.07 to.59, with a median of.28 [interquartile range (IQR):.22 to.34]. When items were residualized, the correlations ranged from.01 to.41, with a median of.12 (IQR:.08 to.16), and 144 (60%) of the correlations were significant. The two vectors of 240 cross-method correlations (for raw item scores and residuals) correlated at.69. These findings suggest that, as a tendency, the items that 22

24 showed higher cross-method agreement were agreed upon not only because they reflected facets (i.e., common variance across particular items), but to a substantial extent because of their distinctive aspects not shared with other items (i.e., nuance-specific item variance). Likewise, items with lower cross-method agreement were not agreed upon regardless of whether we considered their trait-related or unique variance. 2 Stability of Nuances Rank-order stability of nuances over the period of about 5 years was estimated based on the 400 members of the longitudinal sample, treating twin pair members as independent individuals. As in analyses of cross-method agreement, items were residualized for their respective facet scores (with the item itself excluded form it) as well as all other facet scores. In self-ratings, the correlations of raw item scores ranged from.22 to.75 with a median of.53 (IQR:.48 to.58). The correlations ranged from.12 to.61 (Mdn =.34; IQR:.29 to.39) for item residuals, with 232 (97%) being significant. These values are consistent with the variance components of NEO Inventory items estimated by McCrae (2015). Thus, residualizing the item scores for facet variance reduced the median rank-order stability by 36%, but there remained a non-trivial level of stability. In informant ratings, stability was confounded with inter-rater agreement, because the informants did not always overlap at the two time points; 2 Because of the possible effects of dependency in the pooled data, all analyses here, as well as similar analyses elsewhere, were repeated with only one twin from each pair. Results were similar to those reported. Analyses on the predictive validity and demographic associations of item residuals were carried out only on the pooled sample, because maximum power was required. 23

25 correlations were therefore predictably lower than in self-reports. Raw item correlations ranged from.14 to.60 with a median of.37 (IQR:.31 to.44), and residual item correlations ranged from -.01 to.48 (Mdn =.15; IQR:.11 to.22), with 85 (35%) being significant. We also calculated these estimates for combined self- and informant-ratings: the median agreement estimates were.51 and.27, respectively for raw and residualized item scores (193, or 80%, were statistically significant for the latter). The items that had highest rank-order stability when unresidualized tended to be the same items showing highest rank-order stability when residualized for facet variance: Across 240 items, the vectors of rank-order stabilities for raw and residualized items correlated at.66 and.64, respectively, for selfreports and informant ratings (the correlation was.63 for the combined ratings). There was also a substantial level of consistency in rank-order stabilities of items across self-reports and informant ratings, with the correlations across 240 items being.70 for raw item score-based and.59 for item residual-based rank-order stabilities. To the extent that method variance was stable across time, these estimates were inflated by method variance, but they were also attenuated by measurement error. In order to address this possibility, we estimated the stability of true scores by dividing cross-lagged, cross-method correlations by concurrent, cross-method correlations. This resulted in rank-order stabilities of raw item scores ranging from.54 to with a median of.89 (IQR:.81 to.95). For item residuals, the estimated median rank-order stability was.77 (IQR:.58 to 1.00+). These estimates suggest that nuances demonstrate trait-like stability. 24

26 Genetic and Environmental Variance in Nuances Based on the data from 690 twin pairs, we fitted three models (Figure 1) to raw scores and residuals of each of the 240 items, separately for self- and informant ratings. The ACE model specified latent factors representing additive genetic influences (A), shared environmental influences (C) and unique environmental influences (E). The ADE model specified latent factors representing additive genetic influences (A), nonadditive genetic influences due to dominance effects within gene loci (D) and unique environmental influences (E). Finally, the ABE model specified nonadditive genetic (epistatic) influences due to dominance effects between gene loci (B) instead of genetic dominance effects within gene loci. For each item analysis, the three models were compared in terms of comparative model fit based on Akaike Information Criterion (AIC) with the smallest value indicating the best fitting model. We were first interested in the degree to which the preference for any of the three models was replicable across methods. Assuming that the variance captured in self- and informant-reports reflects to some degree the same substantive variance of personality characteristics, we expected to find a substantial degree of convergence. This would have supported the robustness of preferring one model over the others. However, the convergence was modest. In raw item scores, the same type of model was preferred for 97 (40%) items, whereas for item residuals the preference overlapped for 92 (38%) items. Furthermore, the modest convergence could not be ascribed to difficulties between differentiating between ABE and ADE models. If the latter were collapsed into one category of non-ace models and opposed to the category of ACE models, the overlap in preference across rating types was still poor, 25

27 with the same category being preferred for 125 (52%) and 112 (47%) items, respectively for raw scores and residuals. Clearly, preferring one model over another was very inconsistent across methods, as the observed level of convergence could be expected by chance. This prompted us to select only one type of model for further analyses. For the selection of the best model type, we compared intra-class correlations (ICC) in MZ (ICCmz) and DZ twins (ICCdz). Because the ICCdz -s were generally about half the size of ICCmz-s, we opted for, and will only present results of, the ACE models. In raw item scores, the heritability estimates (a²) varied from.00 to.50 (Mdn =.26; IQR:.21 to.33) for self-ratings and from.00 to.48 (Mdn =.19; IQR:.11 to.25) for informant ratings. In item residuals, a² estimates varied from.00 to.37 (Mdn =.14; IQR:.06 to.19) for self-ratings and from.00 to.27 (Mdn =.04; IQR:.00 to.09) for informant ratings. In combined self- and informant-ratings, the median a² estimates varied from.00 to.56 (Mdn =.32; IQR:.24 to.37) and.00 to.41 (Mdn =.13; IQR:.05 to.19), respectively for raw and residualized item scores. Therefore, residualizing the items for facet variance reduced the median a² estimate by 50%, 79% and 59%, for self-reports, informant ratings and combined ratings, respectively. The items that showed higher levels of heritability unresidualized for trait variance tended to be the same that showed higher levels of heritability after residualization. Across the 240 items, raw item-based and residuals-based a² estimates correlated at.61,.54 and.62, respectively for self-reports, informant ratings and combined ratings. The heritability estimates tended to be higher for self-rated than for informant-rated item scores both before and after residualizing for facet variance. Nevertheless, in both types of ratings a nonnegligible proportion of item variance could be ascribed to genetic influences. For example, a² estimates for raw item scores and residuals were significant at p <.0002 for 150 (63%) and 116 (48%) 26

28 items, respectively, based on self-ratings; the corresponding numbers were 121 (50%) and 48 (20%) for informant ratings, and 169 (70%) and 121 (50%) for combined self- and informant-ratings. Heritability estimates for the 240 self-rated items correlated with heritability estimates for the 240 informant-rated items at.36 and.32, respectively for raw item scores and residuals. The items showing significant a² estimates overlapped across methods for 80 (33%) and 24 (10%) items, respectively for raw score- and residual-based estimates. Overall, thus, the residual variances of a tenth of the items showed highly significant evidence for genetic influences simultaneously in self- and informant-reports, suggesting these items consistently reflected some unique signal beyond their role in the measurements of any NEO-PI-R facet (and thereby any FFM domain). In raw item scores, the c² estimates (i.e., variance component due to environmental effects shared by twins) varied from.00 to.37 (Mdn =.00; IQR:.00 to.08) for self-ratings and from.00 to.31 (Mdn =.00; IQR:.00 to.08) for informant ratings. In item residuals, c² estimates varied from.00 to.25 (Mdn =.00; IQR:.00 to.07) for self-ratings, and from.00 to.19 (Mdn =.00; IQR:.00 to.06) for informant ratings; 27 estimates were statistically significant in both cases. In combined self- and informantratings, the median c² estimates varied from.00 to.34 (Mdn =.00; IQR:.00 to.09) and.00 to.22 (Mdn =.00; IQR:.00 to.08; 30 estimates were significant), respectively for raw and residualized item scores. In summary, shared environmental influences were generally small and residualizing items for facet variance had only a small effect on the estimates; if anything, the number of items with significant shared environmental influences increased. Across the 240 items, raw item-based and residuals-based c² estimates correlated at.64,.38 and.66, respectively for self-reports, informant ratings and combined ratings. However, c² estimates for the 240 self-rated items correlated only modestly with c² estimates 27

29 for the 240 informant-rated items (r =.26 and.08, respectively for raw item scores and residuals). Furthermore, the items showing significant c² estimates overlapped across the two rating types for only one and four items, respectively for raw score- and residuals-based estimates. These data provided little consistent evidence of shared environmental effects on nuances. 3 Genetic and Environmental Variance in Nuances after Controlling for Method Effects Estimates of nuance heritability were higher for self-reports than for informant ratings. However, raters of different members of a twin pair were independent, but the twins themselves were not. If rater biases, such as acquiescence, were heritable or influenced by shared (family) experience, twin resemblance may have been inflated in self-reports by shared method artifacts (see Riemann & Kandler, 2010, or Nelling, Kandler, & Riemann, 2015, for a detailed discussion). We therefore created an index of acquiescence from NEO-PI-R items (see Supplement 1) and examined the extent to which genetic and shared environmental effects contribute to acquiescence in self-reports. We found the effects to be non-negligible (a² =.39, c² =.14; see Supplement 1 for details), so we conducted additional analyses of heritability in self-reports controlling for acquiescence (see Supplement 1). This, however, resulted in only minor effects on the median heritability estimates for item residuals (across 3 In order to rule out the possibility that the observed and arguably small a² and c² effects were flukes, we re-ran the analyses on self-reported item residuals 240 times (once for each item) such that the residual scores were reshuffled across individuals and thereby any resulting a² and c² estimates reflected random noise. Only three a² estimates and one c² estimates were significant at p < The corresponding numbers of significant findings had been substantially higher (respectively, 116 and 27) with real data, suggesting that they are meaningful. 28

30 240 items, Mdn a² =.13; IQR:.07 to.19). Of course, other shared method artifacts, such as social desirability, might also account for the higher values in self-reports. In order to circumvent the problem of artifacts, we estimated the genetic and environmental effects on item true scores by employing a procedure similar to that used for estimating the stability of true scores. By calculating ICCdz-s and ICCmz-s across methods, we first assessed the similarity of MZ and DZ twins in a way that eliminated shared method variance (e.g., twin A's self-ratings was correlated with twin B's informant ratings and vice versa; we averaged the ICCs across twincombinations). We then disattenuated these cross-method, cross-twin correlations for imperfect agreement across methods and for random error by dividing them by cross-method, within-twin ICCs (e.g., twin A s self-ratings correlated with twin A s informant ratings and the same for twin B; we used the average correlations across twin-combinations). These estimated twins true score similarity coefficients are designated ICCTRUE-MZ and ICCTRUE-DZ. As the next step, we decomposed the true score variance in raw item scores and residuals into estimates of heritable (a 2 ) and shared environmental (c 2 ) variance, using the classic formulas according to which a 2 = 2*(ICCTRUE-MZ ICCTRUE-DZ) and c 2 = 2*ICCTRUE-DZ ICCTRUE-MZ. Because the ICCs had been derived via a series of calculations, the resulting a 2 and c 2 values were often outside their natural boundaries (i.e., lower than zero or above one); in these cases, the values were set to zero or one. For item raw scores, a 2 estimates varied from.00 to 1.00 with a median of.68 (IQR:.36 to.97). The corresponding c 2 estimates varied from.00 to 1.00 with a median of.00 (IQR:.00 to.26). For item residuals, a 2 estimates ranged from.00 to 1.00 with a median of.52 (IQR:.00 to 1.00). The c 2 estimates varied from.00 to 1.00 with a median of.00 (IQR:.00 to.75). Based on these calculations, thus, 29